feat: support Kubernetes Coscheduling Plugin for PodGroup

[lin121291]

What type of PR is this?

/kind feature

What this PR does / why we need it:

Fixes two behaviors in HAMi that break when used with the Kubernetes Coscheduling Plugin for PodGroup:

1. Concurrent Bind failures — When Coscheduling releases a PodGroup, kube-scheduler binds all member pods near-simultaneously. Multiple /bind calls targeting the same node race on LockNode and fail immediately. LockNode now retries with backoff within a configurable timeout window (--node-lock-retry-timeout, default 30s).

2. Phantom GPU occupancy — When Coscheduling denies a PodGroup, pods return to Pending with stale HAMi annotations. onAddPod re-adds them to podManager cache, creating ghost GPU reservations. Pods in Pending state with stale annotations are now evicted from cache instead.

A third fix cleans up podManager cache in the Bind failure path, which previously leaked ghost entries on failed binds.

Which issue(s) this PR fixes:

Fixes #1832

Does this PR introduce a user-facing change?:

Yes. New flag: --node-lock-retry-timeout (default: 30s).

[hami-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: lin121291
Once this PR has been reviewed and has the lgtm label, please assign wawa0210 for approval. For more information see the Kubernetes Code Review Process.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[gemini-code-assist]

Code Review

This pull request introduces a retry mechanism for node lock acquisition to better handle concurrent bind operations, such as those occurring with coscheduling. It adds a configurable node-lock-retry-timeout flag and environment variable, implements an idempotent lock check, and adds logic to evict pods with stale annotations from the scheduler cache. Feedback indicates that the new eviction logic in onAddPod is too aggressive and may incorrectly remove pods currently in the allocation phase. Additionally, it is recommended to replace the fixed 1-second polling interval in the lock retry logic with an exponential backoff strategy to prevent excessive load on the Kubernetes API server.

[codecov]

Codecov Report

❌ Patch coverage is 61.76471% with 13 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pkg/util/nodelock/nodelock.go	72.41%	7 Missing and 1 partial ⚠️
pkg/scheduler/scheduler.go	0.00%	4 Missing and 1 partial ⚠️

Flag	Coverage Δ
unittests	58.51% <61.76%> (+0.45%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
pkg/scheduler/config/config.go	78.16% <ø> (ø)
pkg/scheduler/scheduler.go	48.66% <0.00%> (-0.86%)	⬇️
pkg/util/nodelock/nodelock.go	63.92% <72.41%> (+1.04%)	⬆️

... and 9 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[mesutoezdil]

Note: As of today, this issue/PR has been flagged under our new activity policy. If there is no response or update by July 18, 2026, it will be closed. It can be reopened at any time if still relevant.